RNA Structural Alignment with Conditional Random Fields

Computationally identifying non-coding RNA regions on the genome has much attention to be investigated. However, it is essentially harder than gene-finding problems for protein-coding regions because non-coding RNA sequences do not have a strong statistical signals. Since comparative sequence analysis is effective for non-coding RNA detection, efficient computational methods are expected for structural alignment of RNA sequences. [5] has proposed one of such algorithms, called pair hidden Markov models on tree structures (PHMMTSs), which can calculate a structural alignment of a binary tree and a sequence, and has applied PHMMTSs to aligning RNA secondary structures, that is, pairwise alignment to align an unfolded RNA sequence into an RNA sequence of known secondary structure. To calculate structural alignments of RNA sequences requires some parameters, the substitution probability of base pairs and the state transition probability, which have much affect on the performance of structural alignments. There are some related works to estimating the parameters for aligning RNA secondary structures. For example, [3] have proposed a ribosomal RNA substitution matrix, called RIBOSUM, which is based on an analogous method to the BLOSUM matrices. However, since the RIBOSUM matrix is based on the maximum likelihood estimation by relative frequencies of RNA mutations, it requires a large number of high-quality structure-annotated alignments to avoid overfitting. Therefore, more effective methods for estimating the parameters for RNA structural alignment should be developed.